Machine learning-basic unsupervised methods (Cluster analysis methods, t-SNE)

Abstract

Understanding how trained deep neural networks achieve their inferred results is challenging but important for relating how patterns in the input data affect other patterns in the output results. We present a visual analytics approach to this problem that consists of two mappings. The so-called forward mapping shows the relative impact of user-selected input patterns to all elements of the output. The backward mapping shows the relative impact of all input elements to user-selected patterns in the output. Our approach is generically applicable to any regressor mapping between two multidimensional real-valued spaces (input to output), is simple to implement, and requires no specific knowledge of the regressor's internals. We demonstrate our method for two applications using image data-a MRI T1-to-T2 generator and a MRI-to-pseudo-CT generator.